we introduce DePO (Decoupled Policy Optimization) to enable effective reinforcement learning within this hybrid space. DePO decomposes the policy gradient objective， a framework that seamlessly interleaves discrete text generation with continuous visual latent representations. Specifically， alongside an exact closed-form von Mises-Fisher (vMF) KL regularizer. Extensive experiments demonstrate that HyLaR outperforms standard MLLMs and state-of-the-art latent reasoning approaches across fine-grained perception and general multimodal understanding benchmarks. Code is available at https://github.com/EthenCheng/HyLaR. ， Chain-of-Thought (CoT) reasoning significantly elevates the complex problem-solving capabilities of multimodal large language models (MLLMs). However， adapting CoT to vision typically discretizes signals to fit LLM inputs， following an initial cold-start supervised fine-tuning (SFT)， we propose HyLaR (Hybrid Latent Reasoning)， applying independent trust-region constraints to the textual and latent components。
confining reasoning to predefined operations. Although recent latent reasoning paradigms internalize visual states to overcome these limitations， they introduce a rigid bottleneck， causing early semantic collapse and discarding fine-grained details. While external tools can mitigate this， optimizing the resulting hybrid discrete-continuous action space remains challenging. In this work，。